Connect PortMem to Claude or ChatGPT and the assistant can work from your own source documents and check a draft against them before you rely on it. Every verdict comes with the passage that decides it. The decisions that matter stay with you, in the PortMem console.
https://mcp.portmem.com/mcp
Claude. Open Customize, then Connectors, then Add and Add custom connector. Paste the address and choose Add. Claude opens PortMem: sign in and approve.
ChatGPT. Open Plugins, then Add and Create custom MCP server. Paste the address with OAuth sign-in and create it. ChatGPT opens PortMem: sign in and approve.
You need a PortMem account. Each person connects with their own account and the assistant only sees that person's workspace. You can see and disconnect your assistants at any time in the console under Settings, Assistants; access stops at once.
List your documents. Your own documents, those shared with you, and public reference documents.
Survey a document. The parties, key dates, amounts and material provisions of one document, read from the whole text. The first survey of a long document can take a few minutes; the assistant is told to search meanwhile, and the result is kept for next time.
Search. The passages that answer a question, with the document each one comes from.
Add a document. Text you give the assistant becomes a private document in your workspace. The assistant only adds new documents. Replacing a document you already have is done in the console.
Check a draft. A review of a memo, summary or report against your documents. Each claim comes back supported, contradicted or unsupported with its evidence, along with the material provisions the draft leaves out. Long reviews run in the background and the assistant checks on them.
Open past reviews. Your earlier reviews and their records, with a link to each in the console.
An assistant cannot accept a flagged claim, publish a record or change who can see a document. Those are your decisions, made in the PortMem console, where they become part of the signed record. When one is needed, the assistant gives you the link.
Read only. When you approve an assistant you can tick Read only. It can then read and search your documents and open your reviews, but cannot add a document or start a review. Settings, Assistants shows which connections are read only.
Every tool works only inside the signed-in person's own workspace. No tool takes a payment, sends a message, publishes, shares or deletes anything. Read tools change nothing. The two write tools only add something new, a document or a review record, and neither is a sensitive write. A read-only connection has the read tools only.
list_sources (read). No inputs. The names of the documents the person can use: their own, those shared with them, and public reference documents.
search (read). Inputs: query; optional source or sources (document names), max_results (default 8), expand (adds the surrounding text). Returns matching passages, each with the document it comes from.
survey_source (read). Inputs: source; optional refresh. Returns the parties, key dates, amounts and material provisions of one document, read from the whole text. The first survey of a long document can answer that it is still preparing; the result is then kept, so asking again is instant. The document itself is not changed.
upload_source (write). Inputs: source_name, text. Adds the text as a new private document. If a document with that name already exists nothing changes and the tool says so; sending the same text again changes nothing.
start_review (write). Inputs: document (the draft); optional source, sources or source_groups to check against, max_claims (default 30). Starts a review in the background and returns a job_id. When it finishes, a new review record is added to the person's ledger. It accepts, publishes and shares nothing.
review_status (read). Input: job_id. Returns the status (queued, running, done, error or cancelled) and progress; once done, every claim with its verdict and evidence, the provisions the draft leaves out, and the console link for the record.
past_reviews (read). Optional limit (default 25). The person's earlier review records, newest first.
get_decision (read). Input: request_id. One review record: its claims, verdicts, whether it is signed, and its console link.
list_signoffs (read). Input: request_id. The sign-offs people recorded on that record in the console.
usage (read). No inputs. The workspace's queries and estimated cost this month.
capabilities (read). No inputs. The retrieval features this deployment has switched on.
Errors and limits. A tool that cannot do what was asked returns a result with an error message saying why, for example an unknown document name, an existing document name, or a review that is still running. An expired or disconnected sign-in gets HTTP 401 and the assistant signs in again. A review can take a few minutes, so the assistant polls review_status. Usage counts toward the workspace like any other use of PortMem.
PortMem offers three in the assistant's menu: Check a draft against my documents, Key facts of a document, and Find something in my documents. Or just ask, for example:
"Check this draft against my credit agreement and tell me what it gets wrong."
"What are the key dates and amounts in the bond terms?"
"Find what my documents say about events of default."
The assistant receives only what the PortMem tools it calls return, such as search passages, survey facts and review results; your assistant's provider then handles that under your agreement with them. Government ID numbers (CPF, CNPJ, SSN) in those results are masked. PortMem stores the connection (the assistant's name and when you connected) and its sign-in tokens only as one-way hashes. Usage through an assistant counts toward your workspace like any other. The privacy policy has the details.
A remote MCP server over Streamable HTTP at https://mcp.portmem.com/mcp, stateless, serving MCP protocol 2026-07-28 and earlier versions. Sign-in is OAuth 2.1 with dynamic client registration and PKCE. Every tool carries read-only, destructive and open-world hints.
Write to info@portmem.com. If a review looks wrong, send its record identifier (it looks like PM-WSVR-45R4-KX3M) and we can find exactly what ran without asking for your documents. More on the support page.